Andscacs - New version 0.85
Moderators: hgm, Rebel, chrisw
-
- Posts: 154
- Joined: Thu Oct 03, 2013 4:17 pm
Re: Andscacs - New version 0.85
Congratulations on your nice progress with evaluation improvements. For some reason every time I have tried to add a new evaluation term it is ±0 elo or worse.
Functional programming combines the flexibility and power of abstract mathematics with the intuitive clarity of abstract mathematics.
https://github.com/mAarnos
https://github.com/mAarnos
-
- Posts: 2204
- Joined: Sat Jan 18, 2014 10:24 am
- Location: Andorra
Re: Andscacs - New version 0.85
Sure at some point I will do it. Initially I prefer to improve the endgame phase without tablebases, but as it can be very boring to view such bad evaluated endgames maybe I will do earlier than I had thought, for the good of spectators .overlord wrote:Nice to see another progress of your engine. Do you plan to add support of syzygy tablebases? Andscacs does not understand e.g. rook endings where white has a+c pawns and it is known draw, but evaluation is +2,45 for a long time.
I just downloaded your last Hakkapeliitta version. I can propose a lot of things that should workBloodbane wrote:Congratulations on your nice progress with evaluation improvements. For some reason every time I have tried to add a new evaluation term it is ±0 elo or worse.
For example in passed pawns:
* bonus for joined passed pawns.
* penalization for passed not controlled.
* penalization for king near passed pawn.
* bonus for own king near passed pawn.
I have even more things only related to passed pawns, but with this you should obtain something. For example the first one is quite easy and gives immediately some win. It's near impossible you don't obtain nothing with it. If is not the case, tell me/us and I will try with my own ways of testing, if you want.
Daniel José - http://www.andscacs.com
-
- Posts: 154
- Joined: Thu Oct 03, 2013 4:17 pm
Re: Andscacs - New version 0.85
The last two I actually have, at least in 3.22, and they gave maybe 10 elo. The relevant lines of code are 360-363 in eval.cpp where 'from' is the square of the current passed pawn we are evaluating. Forgot to document this for some reason so you probably missed it.cdani wrote:I just downloaded your last Hakkapeliitta version. I can propose a lot of things that should work :-)
For example in passed pawns:
* bonus for joined passed pawns.
* penalization for passed not controlled.
* penalization for king near passed pawn.
* bonus for own king near passed pawn.
I have even more things only related to passed pawns, but with this you should obtain something. For example the first one is quite easy and gives immediately some win. It's near impossible you don't obtain nothing with it. If is not the case, tell me/us and I will try with my own ways of testing, if you want.
Code: Select all
const auto distance1 = mChebyshevDistance[kingLocations[!c]][from + 8 - 16 * c];
const auto distance2 = mChebyshevDistance[kingLocations[c]][from + 8 - 16 * c];
scoreEdForColor += static_cast<int>(std::round((std::sqrt(distance1 + 1) - 1) * 39));
scoreEdForColor -= static_cast<int>(std::round((std::sqrt(distance2 + 1) - 1) * 34));
Functional programming combines the flexibility and power of abstract mathematics with the intuitive clarity of abstract mathematics.
https://github.com/mAarnos
https://github.com/mAarnos
-
- Posts: 1242
- Joined: Sat Jul 05, 2014 7:54 am
- Location: Southwest USA
Re: Andscacs - New version 0.85
Thanks Jose!! Testing underway.....methinks its "tactical pressure"cdani wrote:I published a new version, 0.85.
Changes:
* Solved little bugs.
* Improved evaluation of knights.
* Improved evaluation of pawns.
* Improvements on king safety.
* Improvements on space evaluation.
* Improvements on opening phase.
* Some small speed improvements.
* Fixed ponder mode.
http://www.andscacs.com/andscacs085.zip
Seems not much stronger, may be 15-20 elo, but I expect it to scale better at longer time controls or with more processors. Will see
that may be the "extra push"....but how that codes out....your guess is as good as mine.....close to the mountaintop....
-
- Posts: 2204
- Joined: Sat Jan 18, 2014 10:24 am
- Location: Andorra
Re: Andscacs - New version 0.85
I put a list of detailed changes apart of the retuning, bug solving, and improvements of speed:
* Created an evaluation term about compensation between mobility and material in endgame.
* Removed refutation table which was dependent of the 4 previous moves.
* If at search_root one move returns immediately because was found in hash, force a new analysis at depth+1, with some restrictions.
* A little more preference for some aggressive quiet moves in move ordering.
* Increment penalization for a weak pawn that cannot advance easy and is next to an open column.
* Added a little penalization for passed pawns that the advance square is only controlled by rival.
* New table driven move suggestions (up to 8) related to pawn structure (+- 1000 different ones). Adds something to History at the start of the search. Used +-500.000 games of 2300-2800 human players to decide which moves.
* Improved front of the king pawn structure analysis.
* Created an evaluation term about compensation between mobility and material in endgame.
* Removed refutation table which was dependent of the 4 previous moves.
* If at search_root one move returns immediately because was found in hash, force a new analysis at depth+1, with some restrictions.
* A little more preference for some aggressive quiet moves in move ordering.
* Increment penalization for a weak pawn that cannot advance easy and is next to an open column.
* Added a little penalization for passed pawns that the advance square is only controlled by rival.
* New table driven move suggestions (up to 8) related to pawn structure (+- 1000 different ones). Adds something to History at the start of the search. Used +-500.000 games of 2300-2800 human players to decide which moves.
* Improved front of the king pawn structure analysis.
Daniel José - http://www.andscacs.com
-
- Posts: 4465
- Joined: Fri Apr 21, 2006 4:19 pm
- Location: IASI - the historical capital of MOLDOVA
- Full name: SilvianR